It seems the actual speech recognition happens on a remote server, not in the device itself.